Repository navigation

Let the CUDA allocator manage cuBLAS workspaces - #3625

Open

Vtmpas wants to merge 18 commits into

NVIDIA:mainfrom

Vtmpas:fix/cublas-workspace-per-stream

Open

Let the CUDA allocator manage cuBLAS workspaces#3625
Vtmpas wants to merge 18 commits into
NVIDIA:mainfrom
Vtmpas:fix/cublas-workspace-per-stream

Commits on Oct 5, 2026
- Cache cuBLAS workspace by CUDA stream
  Vtmpas
  committed
  93a80c7
- Let the CUDA allocator own cuBLAS workspaces
  Vtmpas
  committed
  e5b55e9
- Test concurrent router dgrad workspace ownership
  Vtmpas
  committed
  ec5f5f9
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  890476d
Commits on Oct 6, 2026
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  77b3d30
Commits on Oct 7, 2026
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  7f36289
Commits on Oct 8, 2026
Commits on Oct 9, 2026
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  baef6e8
- Merge branch 'main' into fix/cublas-workspace-per-stream
  ksivaman
  authored
  af24336
Commits on Oct 12, 2026
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  527b22f
- Merge branch 'main' into fix/cublas-workspace-per-stream
  Vtmpas
  authored
  d50e4ca
Commits on Oct 20, 2026
- Merge branch 'main' into fix/cublas-workspace-per-stream
  
  ksivaman
  and
  claude
  committed
  85c43ce
- Don't allocate workspace for E5M3 path and cache the workspace size
  ksivaman
  committed
  8f6ed78